post-moveco3-1-topicanalysis.utf8.md

This is the companion website for the manuscript “A decade of movement ecology”, from Joo et al., available as a preprint here [PASTE LINK]. It is composed of a series of posts containing the following:

Dimension	Title	Keywords	Abstract	M&M
Topics			X
Taxonomy	X	X	X
Framework	X	X	X
Devices	X	X	X	X
Software	X	X	X	X
Methods	X	X	X	X

3.1. Topic analysis

Stages of topic analysis.

The topics were not defined a priori. Instead, we fitted Latent Dirichlet Allocation (LDA) models to the abstracts (Blei et al. 2003). LDAs are basically three-level hierarchical Bayesian models for documents (in our case, abstracts). Here we assumed that there were latent or hidden topics behind the abstracts, and that the choice of words in the abstracts were related to the topics the authors were addressing. Thus, an abstract would have been composed of one or more topics, and a topic would have been composed of a mixture of words. The probability of a word appearing in an abstract depended on the topic the abstract was adressing. There are several variations of topic models (Nunez-Mir et al. 2016). Here we used the LDA model with variational EM estimation (Wainwright and Jordan 2008) implemented in the topicmodels package. All the details of the model specification and estimation are in Grün and Hornik 2011.

Schematic representation of the links between words, documents and topics. Each document is a mixture of topics. Each topic is modeled as a distribution of words. Each word comes out of one of these topics. Source of the image: Blei, D.M. 2012. Probabilistic topic models. Communications of the ACM, 55(4), 77-84.

Preprocessing

To improve the quality of our LDA model outputs, we cleaned the data by 1) removing unuseful words for identifying topics (e.g. prepositions and numbers), 2) converting all British English words to American English so they would not be seen as different words, 3) lemmatizing (i.e. extracting the lemma of a word based on its intended meaning, with the aim of grouping words under the same lemma) (Ingason et al. 2008), 4) filtering out words that were only used once in the whole set of abstracts. R packages tidytext (Silge and Robinson 2016), tm and textstem (Rinker 2018) were used in this stage.

[PASTE LINK TO CODES]

Fitting LDAs

The parameter estimates of the LDA model were obtained by running 20 replicates of the models (with the VEM estimation method), and keeping the one with the highest likelihood [RJ: add link to the code]. A key argument in the fitting function is the number of latent topics. The most commonly used criterion to choose a number of topics is the perplexity score or likelihood of a test dataset (De Waal and Barnard 2008). Basically, this quantity measures the degree of uncertainty a language model has when predicting some new text (for this study, a new abstract of a paper). Lower values of the perplexity is good and it means the model is assigning higher probabilities. However, the perplexity score measures predictive capacities, rather than having actual humanly-interpretable latent topics (Chang et al. 2009). In fact, using this score could result into having too many topics; see Griffiths and Steyvers (2004) who analyzed PNAS abstracts and obtained 300 topics. Hence, we decided to fix the number of topics to 15, as a reasonable value that would not be too large than we could not interpret them, or too small that the topics were too general.

[PASTE LINK TO CODES]

Model outputs

From the fitted LDA model, we obtained:

the probabilities of having a word in a document given the presence of a certain topic (denoted by \(\beta\)), and
the probabilities of a topic being referred to in each document (denoted by \(\gamma\)).

The \(\beta\) values were thus a proxy of the importance of a word in a topic. They were used to interpret and label each topic, and to create wordclouds for each topic, where the area occupied by each word was proportional to its \(\beta\) value.

Wordclouds of each topic based on \(\beta\) values.

Since \(\gamma\) indicated the degree of association between an abstract and a topic, we obtained a sample of the 5 most associated abstracts to each topic, to aid the interpretation of the topics. [PASTE LINK TO .CSV]

Based on these outputs, the topics were interpreted as: 1) Social interactions and dispersal, 2) Movement models, 3) Habitat selection, 4) Detection and data, 5) Home-ranging, 6) Aquatic systems, 7) Foraging in marine megafauna, 8) Biomechanics, 9) Acoustic telemetry, 10) Experimental designs, 11) Activity budget, 12) Avian migration, 13) Sports, 14) Human activity patterns, 15) Breeding ecology. For an extended description of these topics, see [PASTE LINK TO PRE-PRINT]

The sum of \(\gamma\) values for each topic served as proxies of the “popularity” of the topic relative to all other topics and were used to rank them.

Measure of popularity of each topic.

Time series of the relative popularity of each topic every year.

To check for consistency, for each topic, we selected the papers that were highly associated with the topic (\(\\gamma \> 0.75\)), and computed the number of times each unique word occurred in the abstracts related to the topic. We divided those values by the total number of words in the topic to get a relative frequency, denoted by \(\\delta\). We then created wordclouds for each topic, where the area occupied by each word was proportional to its \(\\delta\) value. These wordclouds were consistent with the topic wordclouds, and gave complementary information.

Wordclouds of each topic based on most strongly associated abstracts.

Also for consistency, a heatmap of the \(\\gamma\) values also showed that most papers were evidently more associated to one topic and few were splitted into several topics.

Heatmap of \(\gamma\) values per abstract and topic.

All the codes to get the model outputs are in [PASTE LINK]

Blei, David M., Andrew Y. Ng, and Michael I. Jordan. 2003. “Latent Dirichlet Allocation.” Journal of Machine Learning Research 3 (Jan): 993–1022. http://jmlr.csail.mit.edu/papers/v3/blei03a.html.

Chang, Jonathan, Sean Gerrish, Chong Wang, Jordan L Boyd-Graber, and David M Blei. 2009. “Reading Tea Leaves: How Humans Interpret Topic Models.” In Advances in Neural Information Processing Systems, 288–96.

De Waal, Alta, and Etienne Barnard. 2008. “Evaluating Topic Models with Stability.” In, 5221:79–84. Nineteenth Annual Symposium of the Pattern Recognition Association of South Africa (PRASA 2008).

Griffiths, Thomas L., and Mark Steyvers. 2004. “Finding Scientific Topics.” Proceedings of the National Academy of Sciences 101 (suppl 1): 5228–35. https://doi.org/10.1073/pnas.0307752101.

Grün, Bettina, and Kurt Hornik. 2011. “topicmodels: An R Package for Fitting Topic Models.” Journal of Statistical Software 40 (13): 1–30. https://doi.org/10.18637/jss.v040.i13.

Ingason, A. K., S. Helgadóttir, H. Loftsson, and E. Rögnvaldsson. 2008. “A Mixed Method Lemmatization Algorithm Using a Hierarchy of Linguistic Identities (Holi).” Advances in Natural Language Processing 5221: 205–16. https://doi.org/10.1007/978-3-540-85287-2_20.

Nunez‐Mir, Gabriela C., Basil V. Iannone, Bryan C. Pijanowski, Ningning Kong, and Songlin Fei. 2016. “Automated Content Analysis: Addressing the Big Literature Challenge in Ecology and Evolution.” Methods in Ecology and Evolution 7 (11): 1262–72. https://doi.org/10.1111/2041-210X.12602.

Rinker, Tyler W. 2018. textstem: Tools for Stemming and Lemmatizing Text. Buffalo, New York. http://github.com/trinker/textstem.

Silge, Julia, and David Robinson. 2016. “Tidytext: Text Mining and Analysis Using Tidy Data Principles in R.” JOSS 1 (3). https://doi.org/10.21105/joss.00037.

Wainwright, Martin J., and Michael I. Jordan. 2008. “Graphical Models, Exponential Families, and Variational Inference.” Foundations and Trends® in Machine Learning 1 (1–2): 1–305. https://doi.org/10.1561/2200000001.

A decade of movement ecology

3. Data analysis

3.1. Topic analysis

Preprocessing

Fitting LDAs

Model outputs